Last Updated: 16 September 2026
Gemini can now keep working even when you keep talking. That is one of the biggest additions in Gemini 3.8 Live, Google's new voice model for conversations that involve more than a quick question and answer. A task can continue in the background as you speak, which means you can add context or ask something else without waiting for the system to complete every step first.
There is also an Extended Thinking version for more difficult requests. It can spend more time reasoning and give spoken updates as it works through the task. Both models support visual information, outside tools and automatic switching between 97 supported languages. The aim is to make voice interactions useful for longer jobs that involve several actions, rather than limiting them to simple questions and short answers.
Waiting has always been one of the awkward parts of using voice AI. Gemini 3.8 Live is designed to reduce those pauses by letting outside actions continue in the background. You can ask the assistant to do something and keep speaking before that job is finished. The session stays open, giving you room to add information or change part of the request as the work continues.
This could be useful when a task depends on another service. A travel assistant might begin checking available flights and still listen as you explain which times you prefer. A support assistant could start looking up an account and answer another question during the search. The conversation can keep moving because the model does not need every connected action to finish before it responds again.
The same session can also include visual information. A camera can provide context that the assistant uses during the conversation. You might show a screen, object or document and ask for help with what appears there. That gives developers more options for voice tools that need to combine spoken requests with outside software and information coming directly from what the user is looking at.
Extended Thinking adds more reasoning to Google's new voice system and is aimed at tasks that cannot be finished with a quick answer. The model can start working through a request and continue speaking during that process. It can confirm what you asked, explain its progress and move through several steps without making the entire conversation stop until the work is complete.
A booking request shows how this could work. You might ask the assistant to compare several flights and find one that fits your schedule. It can begin checking options, use connected tools and tell you how the search is progressing. If you remember another condition halfway through, you can add it during the same conversation instead of waiting for a final answer first.
This gives voice assistants a larger role in tasks that involve several services. The system can call software tools, process new instructions and continue responding by voice during the same session. Developers decide which services are connected, so each application can behave differently. The model provides the reasoning and conversation layer needed to coordinate those actions without turning every step into a separate interaction.
Known AI problems can still appear during longer tasks. Hallucinations, slower responses and timeouts remain possible, which makes verification important when the result affects a real decision.
This new model can follow conversations that move between 97 supported languages without asking users to change settings manually. That can make a real difference for bilingual speakers because switching languages is often part of normal speech. People may use one language for most of a sentence and another for a specific word or phrase. The system is designed to recognize those changes as they happen.
There are also practical uses in education. A student could ask a question in their first language and then continue practicing in a second one. Language tutors could move between both during an explanation without constantly adjusting the interface. That gives learners more freedom to use the language they know when they get stuck, then return to the language they are trying to practice.
Real conversations are more difficult than simply detecting two clearly separated languages. People mix accents, shorten words and sometimes combine languages inside the same sentence. Those details can confuse voice systems. The feature will be easier to judge once users test it across different countries and speaking styles, especially in situations where the language changes several times during a short conversation.
Google is placing the new Gemini voice models across several services instead of keeping them inside one app. Search Live gets the standard model, and Gemini Live gets Extended Thinking. Some Workspace users can also access the new voice features through Gmail, Docs and Keep. For developers, both models are available through Google AI Studio and the Gemini API for building their own products.
The rollout depends heavily on where you are using Gemini. Search and the Gemini app offer the clearest consumer access, while Workspace features require an eligible subscription. Gemini Enterprise is taking a different route with a private preview for both models. That means two people using Google services may see very different access depending on their plan, product and account type.
The highest score highlighted for Gemini 3.8 Live Extended Thinking is 82.6% on the Artificial Analysis Speech to Speech Quality Index. The model also scored 68.6% on τ Voice and 35.1% on the banking version. Big Bench Audio produced a 97.7 percent result. These tests cover different skills, so no single number gives a complete picture of overall voice performance.
Another result comes from Speech Agent Arena, where the standard Live model reached second place. ServiceNow also tested the new models with EVA Bench, which looks at whether an AI completes a task successfully and keeps the conversation useful. That gives developers another measure beyond speech quality alone, especially for applications where getting the requested job done is just as important as sounding natural.
Benchmarks make models easier to compare because everyone faces the same set of tasks. They still leave out much of what happens in normal use. A speaker may pause unexpectedly, use slang or talk over background noise. Requests can also be vague. Those situations can produce very different results from the clean prompts and fixed scoring systems used in many technical evaluations.
For now, the scores are useful reference points. Broader use will show whether that strong test performance continues across ordinary conversations and less predictable situations.
With this new feature is clear that Google is giving voice a bigger place across its products and making spoken interaction useful for more complicated requests. If the experience works well, users may begin choosing speech for tasks they currently handle through text, menus or several separate apps.
That change will not happen simply because the technology sounds more natural. People need to trust that the assistant understands the request and produces useful results consistently. The first benchmark scores are promising, and the rollout reaches a wide range of users. The next stage is less about what the models can demonstrate and more about whether they remain dependable during ordinary use.
The two models are being added to different Google services. Gemini 3.8 Live is available through Search Live and developer platforms such as AI Studio and the Gemini API. Extended Thinking appears in Gemini Live and selected Workspace products for eligible subscribers. Gmail, Docs and Keep are among the supported apps. Business users can access both versions through a Gemini Enterprise private preview. Availability depends on the service and account type, which means some users may receive the features before others.
Gemini 3.8 Live focuses on faster voice conversations and lower operating costs, making it better suited to applications that handle many interactions. Extended Thinking is built for requests that need more reasoning and several stages before completion. Both models support voice, visual input, connected tools and automatic language switching. Extended Thinking spends more computing time on difficult tasks and can provide spoken progress updates. The better option depends on whether an application needs faster responses or more time for complicated requests.
Yes. Gemini 3.8 Live can still produce incorrect information. Its official model card lists hallucinations as a known limitation, meaning the system may generate details that sound convincing but are wrong. It can also experience slow responses or timeouts. The listed knowledge cutoff is January 2025, although supported products may use Search for newer information. A natural speaking voice does not guarantee accuracy, so important claims should be checked through reliable sources before they influence a serious decision.
Yes. Gemini 3.8 Live supports asynchronous tool use, which allows an outside software action to continue without ending the conversation. A user can make a request and keep speaking while another service is working on it. Developers decide which tools an application can access, so the exact actions will vary between products. This setup could support tasks such as checking reservations, looking up information or working with business systems. The final abilities depend on how each developer connects the model to other services.
Yes. Gemini 3.8 Live can process visual input in addition to speech and text. During a live session, the model can use images or video as context for a question. A user could point a camera at an object, document or screen and ask about what appears there. The model can then include that visual information in its response. Accuracy still depends on image quality and what the system can correctly recognize, so users should check important details when visual interpretation affects a decision.
